AI Framework Improves Dementia Etiology Diagnosis

Accurate dementia etiology diagnosis using artificial intelligence remains a significant challenge due to the complex, overlapping symptoms among various diseases. Furthermore, while combining cross-center samples can scale up dataset size and potentially improve performance, the inherent data heterogeneity across different clinical sites or populations often creates conflicts that conventional multi-task learning paradigms fail to adequately address. To tackle this, researchers propose the Collaborative Meta Knowledge Enhancement (COME) framework for dementia etiology diagnosis. COME explicitly models data heterogeneity by injecting multi-center acquisition semantics, source identifiers, and modality indicators as "heterogeneity-aware embeddings" into a unified Transformer architecture. This approach enables robust scale-up training by explicitly accounting for the diverse characteristics of data from different sources. Additionally, a trust-region constrained optimization scheme is integrated into COME to regularize the model during training, preventing it from learning spurious correlations. Evaluated across seven independent cohorts, COME achieved state-of-the-art in-domain performance with a mean macro-averaged AUC of 85.62%, representing a 4.29-point gain over the strongest baseline. It also demonstrated superior out-of-domain generalization under both cross-center and cross-sequence evaluations. Extensive validation confirmed the alignment between model predictions and established biomarkers (amyloid, tau) and clinical severity, highlighting COME's potential for robust and interpretable dementia diagnostics in real-world settings.